← All transcripts

Top Dev Tool Projects : Open Interpreter, spdlog, Dear ImGui, Cognee, Orca & graphify Transcript, AI Summary & Key Points

ManuAGI - AutoGPT Tutorials · 23 days ago · Science & Technology · 17:29 · EN-US

🧠 AI Summary

I cover 20 open-source developer tools and resources focused on terminal coding agents, local knowledge graphs, parallel agent orchestration, C++ libraries, learning, workflow automation, and reusable agent skills. The projects include tools for running agents locally or remotely, querying codebases, coordinating multiple models, building persistent agent memory, creating interfaces and video productions, and improving developer learning and productivity.

🔑 Key Points

  • Open Interpreter is a terminal coding agent for low-cost models such as GLM, DeepSeek, and Kimi, with MCP, skills, hooks, permissions, provider switching, and local configuration and sessions.
  • Graphify maps project folders into queryable local knowledge graphs using tree-sitter across about 40 languages, with no embeddings or vector store; edges are labeled as extracted or inferred.
  • Orca orchestrates parallel coding agents by giving each one its own Git work tree and providing terminal splits, browser interaction, GitHub and Linear views, diff annotation, SSH work trees, and a scripting CLI.
  • OpenMontage uses an AI coding assistant, YAML pipeline manifests, markdown director skills, and Python tools to automate video research, writing, asset generation, editing, and composition across 12 pipelines.
  • Codebase-memory-mcp indexes codebases into persistent knowledge graphs across 158 languages, stores code relationships in SQLite, and provides 15 tools for search, call tracing, architecture overviews, impact analysis, dead-code detection, and Cypher-style queries.
  • spdlog is a fast C++ logging library supporting multiple sinks, synchronous and asynchronous logging, custom patterns, runtime and compile-time log levels, and a backtrace buffer.
  • Cognee gives AI agents persistent memory through a self-hosted knowledge graph that combines vector embeddings with graph relationships and exposes remember, recall, forget, and improve operations.
  • Codex Orchestration turns one Codex task into a multi-model workflow with roles such as advisor, executor, researcher, writer, designer, and reviewer, while Codex remains the root orchestrator.

✅ Actionable items

  • Install Open Interpreter and start a terminal session by typing I.
  • Run UI Skills through NPX, use its start command to route an agent to relevant skills, and fetch skills by category or name.
  • Install Graphify and map a project to generate a clickable HTML graph, report, and JSON file.
  • Use Build Your Own X to select a technology and follow a step-by-step guide for recreating it from scratch.
  • Fork the Open Source Society University repository and track progress through its self-taught computer science curriculum.
  • Use Orca to run coding agents in isolated Git work trees, compare their results, and merge the best output.
  • Clone OpenMontage and describe a video idea for agent-driven production.
  • Install codebase-memory-mcp and index a project for agent-accessible codebase queries.
  • Browse System Prompts Leaks to study system prompts and contribute additional entries.
  • Add Page Agent through a script tag or NPM, then call execute with a natural-language instruction.
  • Add spdlog to a C++ project and configure its sinks, logging mode, patterns, and log levels.
  • Use Dear ImGui by calling UI functions each frame and connecting its output to a project renderer.
  • Install Cognee and provide agents with persistent memory through its Python API, CLI, local UI, or plugins.
  • Browse Chinese Independent Developer and submit a project through a pull request.
  • Install lfg and use its web UI or phone PWA to control remote agent sessions.
  • Install the Temporal TypeScript SDK and write workflows and activities in TypeScript or JavaScript.
  • Define a goal and completion criteria in loop.js so a worker agent can be checked by a separate verifier.
  • Install Codex Orchestration through the Codex plugin marketplace, select models, and define custom roles.
  • Install self-learning-skills so agents can save reusable procedures and dead ends while excluding secrets.
  • Use Three.js Game Skills with Codex or Claude Code to route game development across gameplay, graphics, UI, debugging, QA, and asset generation.

🤖 AI in practice

Used for

Run terminal-based coding agents using lower-cost models such as GLM, DeepSeek, and Kimmy. 01:03
Give AI coding agents specialized guidance for building user interfaces. 01:53
Turn a project folder into a locally queryable knowledge graph for an AI assistant. 02:36
Run multiple coding agents on the same prompt and compare or merge their results. 05:03
Produce a complete video from a natural-language description using an AI coding assistant. 05:55
Let AI coding agents search and analyze a large codebase through a persistent local knowledge graph. 06:49
Control web-page interfaces with natural-language instructions. 08:33
Give AI agents persistent memory across sessions by storing documents as connected, searchable knowledge. 11:05
Study the hidden instructions that shape AI chat assistants and coding agents. 07:45
Operate coding agents remotely from a phone while they run near repositories and credentials. 12:35
Run an AI agent repeatedly until an independent verifier confirms that its goal is complete. 14:16
Turn one Codex task into a coordinated workflow involving multiple specialized models. 15:01
Capture useful debugging procedures and project knowledge from one coding-agent session for reuse in later sessions. 15:48
Build polished playable 3D browser games through coordinated AI development skills. 16:34

Agents

  • Open Interpreter — Run coding tasks in a terminal using low-cost models. 2 held 01:03
  • Orca — Coordinate multiple coding agents working on the same task. 2 held 05:03
  • OpenMontage — Create videos from a plain-language production brief. 2 held 05:55
  • Page Agent — Control web interfaces and automate forms from natural-language commands. 2 held 08:33
  • loop.js — Continue an AI coding task until an independent verifier accepts the result. 2 held 14:16
  • Codex Orchestration — Coordinate a multi-model workflow from one Codex task. 2 held 15:01
  • Three.js Game Skills — Coordinate AI-assisted development of polished 3D browser games. 2 held 16:34

Advice

  • Use Open Interpreter when you want terminal coding agents while keeping model costs low. for Developers using coding agents on a budget.
    It is tuned for open, low-cost models and supports multiple agent harnesses and local sessions.
  • Route AI interface work through UI Skills instead of relying on generic coding-agent output. for Design engineers and developers building user interfaces.
    The collection provides task-specific guidance and skills for interface work.
  • Map large codebases with graphify before asking an AI assistant to navigate them. for Developers navigating large codebases.
    The local graph makes project structure queryable without sending code away or relying on a vector store.
  • Use Orca when working with a fleet of coding agents and needing to compare their outputs. for Developers running multiple coding agents.
    Isolated Git work trees let several agents work in parallel and make merging the best result easier.
  • Use loop.js for long-running agent tasks that need independent verification. for Developers automating long tasks.
    The loop continues from verifier feedback rather than trusting the worker's own declaration of completion.
  • Install self-learning-skills if you work with coding agents every day. for Developers working with agents daily.
    It preserves reusable procedures and dead ends so later sessions do not repeat the same investigation.

What it could not do

  • graphify uses a model for the semantic pass over documents, PDFs, images, and video, and some graph edges are inferred rather than directly extracted. — The graph labels edges as extracted or inferred so users can distinguish read information from guesses.
  • Page Agent requires the developer to supply a model and API key. — The library removes the need for a browser extension, Python, a headless browser, and a multimodal model, but it does not provide the model or key.
  • Three.js Game Skills require optional external keys for generated 3D models, images, and audio. — Procedural fallbacks are used when those keys are not available.
  • Self-learning-skills does not preserve one-off details and never writes secret values. — It records reusable procedures and where secrets can be found rather than recording the secret values themselves.

🧰 Tools & AI usage

AI is used for

  • Running terminal coding agents — Open Interpreter adapts agent scaffolding to low-cost models and supports MCP, skills, hooks, and permissions.01:03
  • Interface design and implementation — UI Skills gives AI coding agents task-specific guidance for user-interface work.01:53
  • Semantic project analysis — Graphify uses a model for semantic passes over documents, PDFs, images, and video while parsing code locally.02:35
  • Parallel coding-agent execution — Orca runs multiple CLI coding agents in isolated Git work trees for comparison and merging.05:02
  • Video production — OpenMontage uses an AI coding assistant for research, writing, asset generation, editing, and final composition.05:55
  • Codebase understanding — codebase-memory-mcp lets AI coding agents query a persistent graph of code structure and relationships.06:49
  • Web-interface control — Page Agent lets users control a web interface with plain language through an in-page JavaScript agent.08:32
  • Persistent agent memory — Cognee stores agent knowledge in a self-hosted graph combining semantic embeddings and graph relationships.11:03
  • Remote agent operation — lfg provides a private control plane for starting, monitoring, and steering coding agents away from the machine hosting repositories and credentials.12:35
  • Durable multi-step workflows — The Temporal TypeScript SDK supports workflows whose retries, state, and recovery remain correct through crashes, timeouts, and restarts.13:25
  • Verified iterative execution — loop.js runs an agent repeatedly until a separate read-only verification agent accepts the result.14:12
  • Multi-model orchestration — Codex Orchestration assigns specialized roles to several models and coordinates their hand-offs through Codex.14:58
  • Knowledge retention — self-learning-skills teaches coding agents to save reusable procedures and debugging lessons for later sessions.15:47
  • Browser-game development — Three.js Game Skills routes agents across gameplay, graphics, UI, debugging, QA, and asset generation, with evidence-based checks before completion.16:33

🔗 Links mentioned

📄 Transcript

Searchable transcript of Top Dev Tool Projects : Open Interpreter, spdlog, Dear ImGui, Cognee, Orca & graphify — ManuAGI - AutoGPT Tutorials (17:29). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by ManuAGI - AutoGPT Tutorials. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 Every week developers release powerful open-source dev tools, and this weekly dev tool project update video puts them in one place, top trending open-source and best dev tool projects. This week cover terminal coding agents, local knowledge graphs, C++ libraries, learning resources, and agent skills. You'll discover useful and trending developer tools you can start using right away, explained fast and to the point without wasting time.

00:24 Let's get started. >> Before we jump into today's project updates, here's a quick announcement for everyone. We've launched a brand new YouTube channel called AI Agent Studio, dedicated entirely to AI agent projects, tutorials, and tools. So, if you're interested in staying up to date with the latest AI agent open-source projects, learning how to build your own agents, or exploring cutting-edge agent frameworks, make sure to check it out.

00:52 Subscribe now to get weekly videos, in-depth guides, and real-time project breakdowns. The link is right there in the description. Don't miss it. All right, let's get into today's video. >> Project number one, open interpreter, terminal coding agent tuned for low-cost models. Open interpreter is a command-line coding agent built for open low-cost models like GLM, DeepSeek, and Kimmy.

01:15 It runs in your terminal after a one-line install on macOS, Linux, or Windows, and you start a session by typing I. Its main idea is harness emulation. It swaps the agent scaffolding to match what a given model responds to best with options like Claude code, Kimmy CLI, Twin code, and DeepSeek 2E. It runs commands inside native sandboxing, switches providers from the interface, and supports MCP, skills, hooks, and permissions, keeping config and sessions on your machine.

01:43 A built-in QA skill lets models drive real web and native apps. It suits developers who want capable agents on cheaper models. Install it and start a session. Project number two, UI skills, agent skills collection for design engineers. UI skills is a command line tool and skill collection for design engineers working with AI coding agents. It gives an agent focused guidance for interface work.

02:07 So, instead of generic output, the agent follows a skill set matched to the task at hand. Developers run it through NPX with no install step. Use a start command to route their agent to the right skills, browse skills by category such as motion, and fetch a specific one like baseline UI directly. The skills live in the repository as plain files, and a companion site documents them.

02:28 It suits developers who build user interfaces with agent help. Run it and route your agent. Project number three, graphify, turns any project folder into a queryable knowledge graph. Graphify is a command line tool and agent skill that maps a project into a knowledge graph. Your AI assistant can query instead of gripping files. It parses code locally with tree-sitter across about 40 languages, so nothing leaves your machine and uses a model only for the semantic pass over docs, PDFs, images, and video.

02:58 It builds no embeddings and no vector store, just a real graph you traverse with query, path, and explain commands. And every edge is tagged as extracted or inferred, so you know what was read versus guessed. It outputs a clickable HTML graph, a report, and a JSON file, and installs into Claude Code, Cursor, Code X, Gemini CLI, and many more. It suits developers navigating large code bases.

03:21 Install it and map your project. Project number four, build your own X, curated guides for rebuilding technologies from scratch. Build your own X is a curated learning resource that collects step-by-step guides for recreating well-known technologies from scratch. It answers a common problem for developers. Reading about how something works rarely teaches as much as building it.

03:42 The repository is a single organized index grouped by what you build, covering databases, operating systems, programming languages, Git, Docker, web servers, neural networks, regex engines, text editors, emulators, games, and more. Each entry links to an outside tutorial and marks the language used from C and Rust to Python, Go, and JavaScript. It suits developers who learn by building rather than reading.

04:07 Pick a topic and start building. Project number five, Open Source Society University, free self-taught computer science degree curriculum. Open Source Society University is a free self-taught computer science curriculum built entirely from online course materials. It solves a familiar problem for self-learners. Plenty of free courses exist, but no clear order or standard tying them together.

04:29 The curriculum follows undergraduate degree requirements minus general education, and courses are selected only if they run regularly, stay open for enrollment, and match recognized curricular standards. Study moves through intro CS, core CS, advanced CS electives, and a final project spanning programming, math, systems theory, security, applications, and ethics.

04:50 Learners track progress by forking the repository and get support through a Discord community. It suits anyone seeking a rigorous grounding in computing. Fork it and start learning. Project number six, Orca desktop environment for orchestrating parallel coding agents. Orca is an open-source desktop application for running many coding agents side by side.

05:12 It solves the coordination problem that appears once you use more than one agent. Each needs its own isolated space, and their results need comparing in one place. Orca gives every agent its own Git work tree, so you can fan a single prompt across several and merge the best output. It runs any CLI agent, including Claude code, Codex, Cursor, Copilot, and Open Code using your own subscription.

05:34 It adds terminal splits, an embedded browser with a click-to-prompt design mode, native GitHub and linear views, diff annotation, SSH work trees, a scripting CLI, and a phone companion app. Written in TypeScript. It runs on Mac OS, Windows, and Linux. It suits developers working with agent fleets. Download it and run your agents. Project number seven, Open Montage, agentic video production system for coding assistance.

06:00 Open Montage is an open-source agent-driven video production system that turns an AI coding assistant into a full production studio. You describe what you want in plain language and the agent handles research, writing, asset generation, editing, and final composition. It has no code orchestrator. The agent reads YAML pipeline manifests and markdown director skills, then calls Python tools through a scored provider selector.

06:23 12 pipelines cover explainers, cinematic pieces, documentary montages, screen demos, dubbing, and more with approval gates, cost estimates, and post-render checks. It works with Claude Code, Cursor, Co-pilot, Windsurf, and Codex. Runs locally with FFmpeg and Remotion, and needs no paid keys to start. It suits creators automating production. Clone it and describe your idea.

06:47 Project number eight, codebase-memory-mepcp, local knowledge graph server for coding agents. Codebase-memory-mepcp is an MCP server that indexes a codebase into a persistent knowledge graph, so AI coding agents can query structure instead of grepping file by file. It parses code with vendor tree-sitter grammars across 158 languages, adds a lightweight type resolution layer for languages like Python, TypeScript, Go, Java, and Rust, and stores functions, classes, call chains, HTTP routes, and cross-service links in

07:19 SQLite. 15 tools cover search, call tracing, architecture overviews, impact analysis, dead code detection, and Cypher-style queries. It ships as a single static binary for Mac OS, Linux, and Windows, runs entirely on your machine, and configures detected agents automatically. It suits developers navigating large repositories. Install it and index your project.

07:42 Project number nine, system prompts leaks, archive of extracted AI assistant system prompts. System prompts leaks is an open reference archive of system prompts captured verbatim from AI chat assistants and coding agents. It addresses a gap most users never see. Every chatbot receives a set of hidden instructions before the first message, and those rules shape its tone, refusals, and tool use.

08:04 The repository organizes files by vendor with folders for Anthropic, OpenAI, Google, XAI, Microsoft, Perplexity, Cursor, Mistral, DeepSeek, and others, covering chat models, coding agents, sub agents, injected reminders, and tool instructions. Everything is plain markdown, kept under a public domain license, and open to contributions. It suits developers and researchers studying prompt design.

08:29 Browse it and study the prompts. Project number 10, page agent, in-page JavaScript agent controlling web interfaces. Page agent is a JavaScript library that puts a GUI agent inside your own web page, so users control the interface with plain language. It removes the usual setup burden, no browser extension, no Python, no headless browser, just in-page script running where the app already lives.

08:52 It reads and acts on the DOM as text rather than screenshots, so it needs no multimodal model or special permissions, and you supply your own model and API key. Developers add it through a single script tag, or install it from NPM, then call execute with an instruction. An optional Chrome extension and MCP server extended across tabs. It suits teams adding co-pilots or form automation.

09:16 Install it and start commanding. Project number 11, SPD log, fast C++ logging library with many sinks. SPD log is a C++ logging library built for speed and simple integration. It handles a need every application shares, recording what happened, where, and when, without slowing the program down. It formats messages through the FMT library and writes to many targets, including rotating files, daily files, colored console output, syslog, the Windows event log, and cute widgets.

09:46 And you can add your own sync. It supports single and multi-threaded loggers, an optional asynchronous mode, custom patterns, runtime and compile time log levels, and a backtrace buffer that holds debug messages until an error occurs. Available header only or compiled, it builds with C make and runs on Linux, Windows, macOS, and Android. It suits C++ developers instrumenting their applications.

10:11 Add it and start logging. Project number 12, Dear ImGui, immediate-mode C++ GUI library for tools. Dear ImGui is a self-contained graphical user interface library for C++ built for tooling rather than end-user apps. It solves a common pain in traditional interfaces, keeping UI state in sync with program state, which invites bugs. In the immediate-mode approach, you simply call functions each frame, and the interface reflects your live data with almost no stored UI state.

10:38 The library outputs vertex buffers and small draw call batches that your own renderer draws, so it never touches the GPU itself. Core files drop into any project without a special build, and bundled backends cover DirectX, OpenGL, Vulkan, Metal, GLOWF, SDL, Win32, and more. It suits engine and real-time developers building debug tools. Add it and start building tools.

11:02 Project number 13, CogniA, open-source memory platform for AI agents. Cogni is an open-source memory platform that gives AI agents lasting recall across sessions. It solves the problem of agents forgetting everything once a conversation ends. You feed in data in any format, and Cogni builds a self-hosted knowledge graph, combining vector embeddings with graph relationships, so documents are searchable by meaning and connected by structure.

11:27 The Python API centers on four operations: remember, recall, forget, and improve with a CLI and local UI alongside. It runs the whole memory layer on a single Postgres instance or swaps in Neo4j, Redis, or other backends and offers Rust and TypeScript clients plus MCP and Claude Code plugins. It suits developers building persistent agents. Install it and give agents memory.

11:51 Project number 14, Chinese Independent Developer, community list of Chinese indie developer projects. Chinese Independent Developer is a community-maintained directory of projects built by independent developers in China. It answers a simple question that is otherwise hard to explore. What are solo builders actually making and where can you find their work in one place?

12:11 The repository is plain markdown, so entries stay easy to read, search, and copy. Alongside the main list, it keeps a separate list for indie games and a programmer-focused edition for developer tools and technical projects. Anyone can add a project through a pull request and contributing guidelines explain the format. It suits indie developers seeking ideas, peers, and visibility for their own work.

12:31 Browse it and add yours. Project number 15, LF Dre, private control plane for remote coding agents. LF Dre is a self-hosted server that turns a Linux box or macOS workstation into a private control plane for AI coding agents. It solves a practical limit. Agents run best where your repositories and credentials live, but you are not always at that machine.

12:55 LF Dre starts each session inside a long-lived tmux session, streams the transcript to a web UI, and lets you answer permission prompts or steer work from a phone through an installable PWA. It supports Claude Code, Codex, and Open Code, binds to loopback by default, and is meant to be reached privately over Tailscale rather than the open internet. Built with Bun and TypeScript, it installs through one setup script.

13:17 It suits developers running agents remotely. Install it and start a session. Project number 16, Temporal TypeScript SDK. Author durable workflows in TypeScript. Temporal TypeScript SDK is the official framework for writing workflows and activities against Temporal, a distributed orchestration engine for long-running business logic. It addresses a hard problem in distributed systems, keeping asynchronous multi-step processes correct when services crash, timeout, or restart.

13:45 Developers write ordinary TypeScript or JavaScript, and the engine handles retries, state, and recovery. The repository is a monorepo of NPM packages split by role with client packages for starting and querying workflows and worker packages for executing them. Worker features rely on node specific capabilities like worker threads, the VM module, and async hooks, and run on node 20, 22, and 24.

14:05 It suits back-end developers building resilient services. Install it and write a workflow. Project number 117, loop.js, framework running agent loops until independently verified. Loop.js is a TypeScript framework and CLI for running an AI agent in a loop until the work is truly finished. It targets the weak point of naive agent loops, the worker declaring itself done.

14:29 You state a goal and what done means, and each round the worker builds with fresh context and memory read back from disk, writes a handoff note, and a separate read-only verify agent judges the result. Only that verdict settles the loop. Rounds, dollars, and timeouts act as guards with a not-yet reason feeding the next round. State lives on disk, so runs resume after crashes, and loop cron schedules entries through your OS scheduler or modal.

14:54 It suits developers automating long tasks. Install it and define a goal. Project number 18, Codex orchestration, plugin turning one Codex task into multi-model workflows. Codex orchestration is a plugin that turns a single Codex task into a workflow across several models. It addresses a limit of one model sessions. The same model plans, builds, and checks its own work.

15:17 Here you bring in other models and assign roles such as advisor, executor, researcher, writer, designer, or reviewer. Then set the order they work in. Codex stays the root orchestrator, handling hand-offs and returning one verified result. And executors can build independent parts in parallel. You install it through the Codex plugin marketplace, run a setup command to pick models, and define custom roles as files in your project or home directory.

15:40 It suits developers running Codex on larger tasks. Install it and design your workflow. Project number 19, self-learning skills. Meta skill teaching agents to keep hard-won knowledge. Self-learning skills is a meta skill for AI coding agents that captures what a session learned, so the next one starts informed. It targets a daily waste, hard debugging, an odd command, or a project fact gets figured out then disappears when the session ends.

16:07 The skill teaches the agent to recognize that moment, write the procedure and the dead ends into a reusable file, and reuse it automatically later. Where it saves depends on the tool, a skill.mb for Claude code, a rule under .cursor/rules for cursor, or agents.mb elsewhere. One-off details are skipped, and secret values are never written, only where to find them.

16:28 It suits developers working with agents daily. Install it and start harvesting. Project number 20, 3js game skills. Agent skills for building polished browser games. 3js game skills is a set of agent skills for building playable polished 3js browser games with Codex or Claude code. It removes the burden of picking the right specialist for each stage.

16:49 You name the director skill, and it routes work across game play systems, graphics, UI, debugging, QA, and asset generation. Each skill is self-contained, carrying its own instructions, references, scripts, and a bundled vite and typescript scaffold for empty projects. Before claiming success, the skills gather evidence through builds, browser runs, screenshots, canvas pixel checks, and mobile viewport passes.

17:14 Optional keys enable generated 3D models, images, and audio with procedural fallbacks otherwise. It suits developers making browser games with agents. Install it and start building. Thanks for watching. See you in the next update.